- Machine Learning
- Product Development
Project Skate is helping to change the face of skateboarding with machine learning.
Interest in skateboarding continues to soar after the sport saw its debut in the Tokyo Olympic Games in 2021. As this sport continues to sky-rocket, so does its fanbase - skaters and spectators both eager to understand the sport more deeply.
The Google team wanted to use this insight to make Skateboarding more accessible to the general public by breaking down the tricks into understandable parts. They partnered with us to create an event-focused technical installation for half-pipe skateboarding. Multiple cameras, Google’s Tensorflow AI platform, and computer vision allowed the audience and skater to watch a data-augmented replay of their session with tricks automatically named and stats generated.
Skating in competitions since the age of eight, Sky Brown won a bronze medal at the 2021 Tokyo Olympics, representing Great Britain
Sky Brown, the 14-year-old female Olympic bronze medallist, helped us to evaluate our system and provided insights into how she does what she does best.
Skateboarding is now an olympic sport, with talented newcomers from around the world competing for the gold. Seeing the stunts that these athletes can pull off is incredible, but it’s also extremely difficult for a beginner skater to replicate these feats with achievable milestones.
The difficulty of explaining the performance isn’t helped by the fact that skateboarding tricks have impenetrable names for the lay-person, like the “Fakie McTwist” or the “Frontside Caballerial”. They are often named after the first person to successfully perform the trick, which preserves the history of the sport but makes the names opaque to the casual observer. Even the breakdown of how a trick is performed is full of jargon: performing a frontside, backside, fakie, blunt, disaster, 180, etc.
We used the open-source Google Tensorflow-based AI system, BlazePose, along with studio-grade cameras and wireless sensors to track the pose of the skater and skateboard at sixty times per second.
We captured hundreds of example tricks on the Southern Hemisphere’s largest half-pipe with some of Australia’s best skateboarders at the Sydney Olympic Park. Hours of footage at 60fps from multiple angles let us see the athletes’ tricks in high detail. We hand labeled each of the tricks with useful information, like their approach side, the number of spins they performed, whether the trick was a grind or aerial, etc.
Project Skate captures Sky's body position at every part of the trick, generating the data behind her movements.
From analysing this footage, we realised that skateboarding tricks are composed of a small set of actions that are strung together in new and interesting ways. So, we developed a program to measure some of the fundamental parts of the skateboard trick, like: how fast the skateboarder was going, how high they reached, how many rotations they performed, where they grabbed the board, and how the board moved in the air or on the rail.
Now that we had the separate parts of the trick, we could combine these systems together and give the trick its name. To make the trick names more friendly and universally accepted, our skateboarders recommended that we should use the tricks’ names from the Tony Hawk Pro Skater video game. We played a lot of that game to get the information we needed.
Project Skate was able to identify Sky's signature tricks, and provide data behind them that she had never seen before.
Three cameras were installed to ensure a view of the skateboarder from all angles.
Using Tensorflow and Blazepose, Project Skate identifies the skateboarder in frame and draws points on parts of their body.
We had a working prototype on-set in San Francisco nine weeks after the initial brief. Sky Brown, 14-year-old Olympic Bronze medallist in the women’s street skateboarding category, was our Beta tester.
Google and Vice, produced a film around this technology that was played at the Google Developers Conference and received much internal and external business interest.